Emotion Recognition with Deep-Belief Networks

For our CS229 project, we studied the problem of reliable computerized emotion recognition in images of human faces. First, we performed a preliminary exploration using SVM classifiers, and then developed an approach based on Deep Belief Nets. Deep Belief Nets, or DBNs, are probabilistic generative models composed of multiple layers of stochastic latent variables, where each “building block” layer is a Restricted Boltzmann Machine (RBM). DBNs have a greedy layer-wise unsupervised learning algorithm as well as a discriminative fine-tuning procedure for optimizing performance on classification tasks. [1]. We trained our classifier on three databases: the Cohn-Kanade Extended Database (CK+) [2], the Japanese Female Facial Expression Database (JAFFE) [3], and the Yale Face Database (YALE) [4]. We tested several different database configurations, image pre-processing settings, and DBN parameters, and obtained test errors as low as 20% on a limited subset of the emotion labels. Finally, we created a real-time system which takes images of a single subject using a computer webcam and classifies the emotion shown by the subject.